Qwen3.8-Max
2.4T MoE, official $2/$6, not in the catalog
Alibaba’s 2026-08-03 Qwen flagship: 2.4T MoE, ~95B active, hosted API id qwen3.8-max, 1M context, published API $2/$6 per 1M plus $0.25 cache. Open weights for Qwen3.8-2.4T-A95B landed ~2026-08-14 on Hugging Face / ModelScope. QCode /models does not list this id.
The specs that matter
MoE size
2.4T total, ~95B active. Alibaba / Qwen official figures.
Context
Hosted API publishes a 1M-token window.
Published API
$2 in / $6 out per 1M; cached input $0.25. That is the public API table, not a QCode tariff.
Hosted API id
This is the official hosted id. It is not on the QCode catalog.
What 3.8-Max is
Qwen3.8-Max is Alibaba’s Qwen flagship announced 2026-08-03. 2.4T MoE, ~95B active; hosted API id qwen3.8-max; 1M context. Official trail: alibabagroup.com document, qwen.ai/blog?id=qwen3.8, @Alibaba_Qwen. First time a Max-class checkpoint shipped with open weights.
Weights are out. Do not equate surfaces.
Around 2026-08-14, Qwen3.8-2.4T-A95B open weights appeared on Hugging Face and ModelScope. Hosted Max may expose vision / tools the raw checkpoint does not. “You can download weights” is not “same capability as the hosted API.”
Timeline
Alibaba announces Qwen3.8-Max. Hosted API id: qwen3.8-max.
Open weights Qwen3.8-2.4T-A95B on Hugging Face / ModelScope. 27B lands the same day.
Official hosting serves qwen3.8-max. QCode /models still lists only qwen3.7-max and qwen3.7-plus.
Confirmed vs treat carefully
Confirmed
Announced 2026-08-03; 2.4T MoE / ~95B active; hosted id qwen3.8-max; 1M context; published API $2/$6 + cache $0.25; Qwen3.8-2.4T-A95B weights ~2026-08-14. Sources: Alibaba document, qwen.ai/blog?id=qwen3.8, @Alibaba_Qwen.
Treat carefully
Do not assume hosted Max equals the raw 2.4T-A95B checkpoint — vision / tools may be hosted-only. Do not treat third-party resale as the official table. Do not assume QCode lists qwen3.8-max. It does not.
Two paths
Official hosted 3.8-Max
Need the new flagship and the official vision / tool surface? Use QwenCloud / the official API, id qwen3.8-max.
Qwen that QCode actually lists
QCode sells qwen3.7-max. 3.8-Max is newer. It is not in the catalog.
How the official surface works
On the hosted API set model=qwen3.8-max. For weights, pull Qwen3.8-2.4T-A95B from Hugging Face / ModelScope. Local inference should not assume hosted vision / tools. This page does not give a QCode call path for 3.8-Max — that id is not listed.
On QCode
QCode /models does not list qwen3.8-max. The Qwen ids on sale are qwen3.7-max and qwen3.7-plus. To call Qwen on QCode today, use qwen3.7-max. 3.8-Max is newer and not in the catalog.
Qwen3.8-Max FAQ
Is 3.8-Max out?
Yes. Alibaba announced it on 2026-08-03.
How big is it?
2.4T MoE, ~95B active. Official figures.
What is the published API price?
$2 in / $6 out per 1M, cache $0.25. Public API table.
Are there open weights?
Yes. Qwen3.8-2.4T-A95B landed ~2026-08-14 on Hugging Face / ModelScope.
Is the checkpoint the hosted Max?
Not necessarily. Hosted Max may have vision / tools the raw checkpoint does not.
Can I call qwen3.8-max on QCode?
No. That id is not listed. For Qwen on QCode, use qwen3.7-max.
Sources
Alibaba Group document; qwen.ai/blog?id=qwen3.8; @Alibaba_Qwen. Prices are the public API table, not a QCode tariff.
Call the Qwen that is already listed
qwen3.8-max is not on QCode. qwen3.7-max is.
Related
Qwen3.7-Max guide
The Qwen flagship QCode actually lists: qwen3.7-max.
Qwen3.8-27B
Dense multimodal you can download. Apache 2.0.
2026 model radar
Which families are listed, which are not.
Not affiliated with Alibaba / Qwen. QCode does not offer qwen3.8-max. Figures are from public official materials; other channels may differ.